[core] LLM.collective_rpc interface and RLHF example (vllm-project#12084

)

Signed-off-by: youkaichao <[email protected]>

.buildkite/test-pipeline.yaml

@@ -126,11 +126,15 @@ steps:
   - tests/distributed
   - tests/spec_decode/e2e/test_integration_dist_tp4
   - tests/compile
+  - examples/offline_inference/rlhf.py
   commands:
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
+  # TODO: create a dedicated test section for multi-GPU example tests
+  # when we have multiple distributed example tests
+  - python3 ../examples/offline_inference/rlhf.py
 
 - label: Metrics, Tracing Test # 10min
   num_gpus: 2 

examples/offline_inference/rlhf.py

@@ -0,0 +1,191 @@
+"""
+a simple demonstration of RLHF with vLLM, inspired by
+the OpenRLHF framework https://github.com/OpenRLHF/OpenRLHF .
+It follows the design that, training processes and inference processes
+are different, and they live on different GPUs.
+Training processes send prompts to inference processes to generate data,
+and also synchronize the weights of the model by broadcasting the weights
+from the training process to the inference process.
+Note that this is a simple demonstration of one training instance and one
+inference instance. In practice, there could be multiple training instances
+and multiple inference instances. For the full implementation, please refer
+to the OpenRLHF framework.
+"""
+import os
+
+import ray
+import torch
+from ray.util.placement_group import placement_group
+from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
+from transformers import AutoModelForCausalLM
+
+from vllm import LLM, SamplingParams, configure_as_vllm_process
+from vllm.utils import get_ip, get_open_port
+from vllm.worker.worker import Worker
+
+
+def stateless_init_process_group(master_address, master_port, rank, world_size,
+                                 device):
+    """
+    vLLM provides `StatelessProcessGroup` to create a process group
+    without considering the global process group in torch.distributed.
+    It is recommended to create `StatelessProcessGroup`, and then initialize
+    the data-plane communication (NCCL) between external (train processes) 
+    and vLLM workers.
+    """
+    from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
+    from vllm.distributed.utils import StatelessProcessGroup
+    pg = StatelessProcessGroup.create(host=master_address,
+                                      port=master_port,
+                                      rank=rank,
+                                      world_size=world_size)
+    pynccl = PyNcclCommunicator(pg, device=device)
+    return pynccl
+
+
+class MyWorker(Worker):
+    """
+    The `MyWorker` class inherits from `Worker` to provide custom functions.
+    For simplicity, we define the `MyWorker` class in this self-contained 
+    script. Normally, we should define the `MyWorker` class in a separate 
+    file and pass the qualified name of the class to the `worker_cls` 
+    parameter.
+    """
+
+    def init_weight_update_group(self, master_address, master_port,
+                                 rank_offset, world_size):
+        from vllm.distributed.parallel_state import get_world_group
+        rank = get_world_group().rank + rank_offset
+        self.model_update_group = stateless_init_process_group(
+            master_address,
+            master_port,
+            rank,
+            world_size,
+            self.device,
+        )
+
+    def update_weight(self, name, dtype, shape):
+        weight = torch.empty(shape, dtype=dtype, device="cuda")
+        self.model_update_group.broadcast(weight,
+                                          src=0,
+                                          stream=torch.cuda.current_stream())
+
+        self.model_runner.model.load_weights(weights=[(name, weight)])
+
+        del weight
+
+    def check_weights_changed(self):
+        """
+        Check if the weights are updated to 0.
+        """
+        weights_updated = True
+        for name, p in self.model_runner.model.named_parameters():
+            weights_updated = weights_updated and torch.allclose(
+                p, torch.zeros_like(p))
+        return weights_updated
+
+
+class MyLLM(LLM):
+
+    def __init__(self, *args, **kwargs):
+        # a hack to make the script work.
+        # stop ray from manipulating CUDA_VISIBLE_DEVICES
+        # at the top-level
+        del os.environ["CUDA_VISIBLE_DEVICES"]
+        super().__init__(*args, **kwargs)
+
+
+"""
+Start the training process, here we use huggingface transformers 
+as an example to hold a model on GPU 0.
+
+It is important for all the processes outside of vLLM to call
+`configure_as_vllm_process` to set some common environment variables
+the same as vLLM workers.
+"""
+configure_as_vllm_process()
+
+train_model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m")
+train_model.to("cuda:0")
+"""
+Start the inference process, here we use vLLM to hold a model on GPU 1 and 
+GPU 2. For the details on how to use ray, please refer to the ray 
+documentation https://docs.ray.io/en/latest/ .
+"""
+os.environ["CUDA_VISIBLE_DEVICES"] = "1,2"
+ray.init()
+
+pg_inference = placement_group([{"GPU": 1, "CPU": 0}] * 2)
+ray.get(pg_inference.ready())
+scheduling_inference = PlacementGroupSchedulingStrategy(
+    placement_group=pg_inference,
+    placement_group_capture_child_tasks=True,
+    placement_group_bundle_index=0,
+)
+"""
+launch the vLLM inference engine.
+here we use `enforce_eager` to reduce the start time.
+"""
+llm = ray.remote(
+    num_cpus=0,
+    num_gpus=0,
+    scheduling_strategy=scheduling_inference,
+)(MyLLM).remote(
+    model="facebook/opt-125m",
+    enforce_eager=True,
+    worker_cls=MyWorker,
+    tensor_parallel_size=2,
+    distributed_executor_backend="ray",
+)
+
+# Generate texts from the prompts.
+prompts = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+
+sampling_params = SamplingParams(temperature=0)
+
+outputs = ray.get(llm.generate.remote(prompts, sampling_params))
+
+for output in outputs:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, "
+          f"Generated text: {generated_text!r}")
+
+# set up the communication between the training process
+# and the inference engine.
+master_address = get_ip()
+master_port = get_open_port()
+
+handle = llm.collective_rpc.remote("init_weight_update_group",
+                                   args=(master_address, master_port, 1, 3))
+model_update_group = stateless_init_process_group(master_address, master_port,
+                                                  0, 3, torch.device("cuda:0"))
+ray.get(handle)
+
+# simulate training, modify the weights of the model.
+for name, p in train_model.named_parameters():
+    p.data.zero_()
+
+# sync weight from the training process to the inference engine.
+for name, p in train_model.named_parameters():
+    handle = llm.collective_rpc.remote("update_weight",
+                                       args=(name, p.dtype, p.shape))
+    model_update_group.broadcast(p, src=0, stream=torch.cuda.current_stream())
+    ray.get(handle)
+
+# check if the weights are updated.
+assert all(ray.get(llm.collective_rpc.remote("check_weights_changed")))
+
+# use the updated model to generate texts, they will be nonsense
+# because the weights are all zeros.
+outputs_updated = ray.get(llm.generate.remote(prompts, sampling_params))
+for output in outputs_updated:
+    prompt = output.prompt
+    generated_text = output.outputs[0].text
+    print(f"Prompt: {prompt!r}, "
+          f"Generated text: {generated_text!r}")

vllm/__init__.py

@@ -17,6 +17,44 @@
 
 from .version import __version__, __version_tuple__
 
+
+def configure_as_vllm_process():
+    """
+    set some common config/environment variables that should be set
+    for all processes created by vllm and all processes
+    that interact with vllm workers.
+    """
+    import os
+
+    import torch
+
+    # see https://github.com/NVIDIA/nccl/issues/1234
+    os.environ['NCCL_CUMEM_ENABLE'] = '0'
+
+    # see https://github.com/vllm-project/vllm/issues/10480
+    os.environ['TORCHINDUCTOR_COMPILE_THREADS'] = '1'
+    # see https://github.com/vllm-project/vllm/issues/10619
+    torch._inductor.config.compile_threads = 1
+
+    from vllm.platforms import current_platform
+
+    if current_platform.is_xpu():
+        # see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158
+        torch._dynamo.config.disable = True
+    elif current_platform.is_hpu():
+        # NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1)
+        # does not support torch.compile
+        # Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for
+        # torch.compile support
+        is_lazy = os.environ.get('PT_HPU_LAZY_MODE', '1') == '1'
+        if is_lazy:
+            torch._dynamo.config.disable = True
+            # NOTE(kzawora) multi-HPU inference with HPUGraphs (lazy-only)
+            # requires enabling lazy collectives
+            # see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501
+            os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true'
+
+
 __all__ = [
     "__version__",
     "__version_tuple__",
@@ -42,4 +80,5 @@
     "AsyncEngineArgs",
     "initialize_ray_cluster",
     "PoolingParams",
+    "configure_as_vllm_process",
 ]

vllm/entrypoints/llm.py

@@ -4,6 +4,7 @@
 from typing import (Any, ClassVar, Dict, List, Optional, Sequence, Tuple, Type,
                     Union, cast, overload)
 
+import cloudpickle
 from tqdm import tqdm
 from typing_extensions import deprecated
 
@@ -186,6 +187,13 @@ def __init__(
         if "disable_log_stats" not in kwargs:
             kwargs["disable_log_stats"] = True
 
+        if "worker_cls" in kwargs:
+            worker_cls = kwargs["worker_cls"]
+            # if the worker_cls is not qualified string name,
+            # we serialize it using cloudpickle to avoid pickling issues
+            if isinstance(worker_cls, type):
+                kwargs["worker_cls"] = cloudpickle.dumps(worker_cls)
+
         if compilation_config is not None:
             if isinstance(compilation_config, (int, dict)):
                 compilation_config_instance = CompilationConfig.from_cli(
@@ -455,6 +463,23 @@ def generate(
         outputs = self._run_engine(use_tqdm=use_tqdm)
         return self.engine_class.validate_outputs(outputs, RequestOutput)
 
+    def collective_rpc(self,
+                       method: str,
+                       timeout: Optional[float] = None,
+                       args: Tuple = (),
+                       kwargs: Optional[Dict] = None) -> List[Any]:
+        """
+        Run a method on all workers, with homogeneous arguments.
+        The main extension point for the LLM entrypoint.
+        Users can provide custom worker class through `worker_cls`
+        argument, and implement new methods in the worker class.
+        Then, users can call the new methods through this API.
+        It is recommended to use this API to only pass control messages,
+        and set up data-plane communication to pass data.
+        """
+        return self.llm_engine.model_executor.collective_rpc(
+            method, timeout, args, kwargs)
+
     def beam_search(
         self,
         prompts: List[Union[TokensPrompt, TextPrompt]],

vllm/plugins/__init__.py

@@ -1,9 +1,6 @@
 import logging
-import os
 from typing import Callable, Dict
 
-import torch
-
 import vllm.envs as envs
 
 logger = logging.getLogger(__name__)
@@ -50,34 +47,6 @@ def load_general_plugins():
     processes. They should be designed in a way that they can be loaded
     multiple times without causing issues.
     """
-
-    # all processes created by vllm will load plugins,
-    # and here we can inject some common environment variables
-    # for all processes.
-
-    # see https://github.com/vllm-project/vllm/issues/10480
-    os.environ['TORCHINDUCTOR_COMPILE_THREADS'] = '1'
-    # see https://github.com/vllm-project/vllm/issues/10619
-    torch._inductor.config.compile_threads = 1
-
-    from vllm.platforms import current_platform
-
-    if current_platform.is_xpu():
-        # see https://github.com/pytorch/pytorch/blob/43c5f59/torch/_dynamo/config.py#L158
-        torch._dynamo.config.disable = True
-    if current_platform.is_hpu():
-        # NOTE(kzawora): PT HPU lazy backend (PT_HPU_LAZY_MODE = 1)
-        # does not support torch.compile
-        # Eager backend (PT_HPU_LAZY_MODE = 0) must be selected for
-        # torch.compile support
-        is_lazy = os.environ.get('PT_HPU_LAZY_MODE', '1') == '1'
-        if is_lazy:
-            torch._dynamo.config.disable = True
-            # NOTE(kzawora) multi-HPU inference with HPUGraphs (lazy-only)
-            # requires enabling lazy collectives
-            # see https://docs.habana.ai/en/latest/PyTorch/Inference_on_PyTorch/Inference_Using_HPU_Graphs.html # noqa: E501
-            os.environ['PT_HPU_ENABLE_LAZY_COLLECTIVES'] = 'true'
-
     global plugins_loaded
     if plugins_loaded:
         return

vllm/worker/worker_base.py

@@ -4,6 +4,7 @@
 from abc import ABC, abstractmethod
 from typing import Any, Dict, List, Optional, Set, Tuple, Type, Union
 
+import cloudpickle
 import torch
 
 from vllm.config import ObservabilityConfig, VllmConfig
@@ -521,14 +522,20 @@ def init_worker(self, all_kwargs: List[Dict[str, Any]]) -> None:
         kwargs = all_kwargs[self.rpc_rank]
         enable_trace_function_call_for_thread(self.vllm_config)
 
-        # see https://github.com/NVIDIA/nccl/issues/1234
-        os.environ['NCCL_CUMEM_ENABLE'] = '0'
+        from vllm import configure_as_vllm_process
+        configure_as_vllm_process()
 
         from vllm.plugins import load_general_plugins
         load_general_plugins()
 
-        worker_class = resolve_obj_by_qualname(
-            self.vllm_config.parallel_config.worker_cls)
+        if isinstance(self.vllm_config.parallel_config.worker_cls, str):
+            worker_class = resolve_obj_by_qualname(
+                self.vllm_config.parallel_config.worker_cls)
+        else:
+            assert isinstance(self.vllm_config.parallel_config.worker_cls,
+                              bytes)
+            worker_class = cloudpickle.loads(
+                self.vllm_config.parallel_config.worker_cls)
         self.worker = worker_class(**kwargs)
         assert self.worker is not None